RecurPost schedules. S4L picks the tone of every reply.

RecurPost's unit of work is a scheduled post drawn from a content library; the tool queues copy you wrote and fires it on a calendar cadence. S4L's unit of work is a reply to somebody else's thread, and the tone of that reply is chosen by a tier-ranked style picker before the prompt is assembled. Concretely: scripts/engagement_styles.py defines eight named styles (critic, storyteller, pattern_recognizer, curious_probe, contrarian, data_point_drop, snarky_oneliner, recommendation) and runs a Postgres query on every invocation to rank them by average upvotes per platform. Nothing in the RecurPost listicle category does this, because their primitive is the scheduled post, not the style-chosen reply.

scripts/engagement_styles.py, function get_dynamic_tiers(platform, context='posting') at lines 162-211. It calls _fetch_style_stats, which runs `SELECT engagement_style, COUNT(*) AS n, AVG(COALESCE(upvotes,0))::float AS avg_up FROM posts WHERE status='active' AND engagement_style IS NOT NULL AND our_content IS NOT NULL AND LENGTH(our_content) >= 30 AND upvotes IS NOT NULL AND platform = %s GROUP BY engagement_style`. Trusted styles (N >= MIN_SAMPLE_SIZE = 5) are sorted by avg_up DESC and split into thirds: top -> dominant, middle -> secondary, bottom -> rare. Anything below the sample threshold (including zero-sample styles) goes to secondary so the model still explores it.

Seven are posting styles, shared across all platforms: critic (point out what's missing or naive, reframe), storyteller (first-person narrative with specific details, lead with failure not success), pattern_recognizer (name the phenomenon, authority through pattern recognition), curious_probe (one specific follow-up question with a 'curious because' clause), contrarian (opposing position backed by experience), data_point_drop (one specific believable metric, numbers must be believable not impressive), and snarky_oneliner (one punchy sentence, NEVER on LinkedIn). The eighth, recommendation, only appears in reply context and is capped at MAX 20% of replies.

It is not a hard selector; it is a directive in the generated prompt. get_styles_prompt emits three headings (PRIMARY, SECONDARY, RARE) with the instruction 'use these ~60% of the time', '~30%', '~10%'. The LLM picks inside those tiers. The empirical split therefore depends on how well the model follows the instruction, which is why the styles list keeps being ranked by live upvotes rather than a fixed mix: any style the model over-fires that under-performs will drop a tier on the next run.

MIN_SAMPLE_SIZE = 5 is the cutoff between 'trusted' and 'explore' styles. A style with fewer than 5 logged posts on this platform is not trustworthy enough to rank on average upvotes, so it goes to secondary (explore) regardless of the noisy avg. Five is low enough that a new platform warms up in a week or two of daily posts, high enough to damp out the first lucky run that lands 400 upvotes on a style that is otherwise a dud.

_fetch_style_stats returns {} and get_dynamic_tiers returns ([], explore, []) where explore = every non-never style. The prompt block ships with no PRIMARY section, no RARE section, and one SECONDARY heading listing every style. The 60/30/10 language is omitted. The model gets told to explore uniformly. Once MIN_SAMPLE_SIZE posts with non-null upvotes land per style, tiering begins to kick in automatically. This is why the first week of S4L's output looks visibly more varied than a tool with a fixed prompt.

Because it performs well on the platforms where it is allowed (large Twitter threads, 500k+ member subreddits). PLATFORM_POLICY drops it for LinkedIn and GitHub because the tone is wrong for professional surfaces, regardless of what the upvote data says. Two separate constraints: tone (never list) and performance (tier sort). They compose; the never list runs after the tier sort, so a style can be top-ranked and still blocked by policy.

recommendation is the only reply-specific style. It appears only when get_styles_prompt is called with context='replying', and it has one job: casually mention a project from config.json. It's capped at MAX 20% of replies because casual project mentions are how social credibility is built on Reddit and LinkedIn, but self-promo past about one-in-five replies is the fastest way to get flagged as a promotional account. The cap is written directly into the generated prompt, not enforced in code, so the model reads it and moderates itself.

Yes. PLATFORM_POLICY defines entries for reddit, twitter, linkedin, github, moltbook. Each has its own 'never' list and its own 'note' (a tone/length hint inserted at the top of the styles prompt). The tier sort runs against the rows in posts that match WHERE platform = %s, so Twitter's dominant styles can be totally different from LinkedIn's on the same day. A style that never gets used on one platform stays in explore on that platform forever without affecting ranking elsewhere.

Because the data showed that replies along the lines of 'this is great', 'had similar results', '100% agree' consistently got the lowest engagement across all three major platforms, even when the surrounding thread was high-signal. The AVOID clause is a hard-coded guard at the end of every generated styles prompt. It's not one of the 8 styles because it isn't a style we want a 60/30/10 mix of; it's a class of output the tool actively refuses.

When a reply is posted, the row inserted into posts carries engagement_style = X, our_content, platform, status='active', and upvotes=NULL. Once the platform returns an upvote count (via the stats collection launchd job, scheduled at 00:45 UTC), that cell populates. From that moment on, the next get_dynamic_tiers call sees the new sample. No batch rebuild, no explicit retraining step. The ranking is just a GROUP BY AVG against whatever rows are in the table right now.

No. It's not the same shape of tool. S4L doesn't have a calendar UI, a content library, bulk upload CSV, an RSS import, or IG DM automation. What it does have is a style-choosing reply pipeline with live upvote ranking, a per-subreddit cooldown floor (see our other guide on the 1 day / 3 day floors), and a set of macOS launchd jobs that run the loop unattended. If your workflow is 'queue evergreen copy, recycle on a cadence', RecurPost fits. If your workflow is 'reply in threads that match my projects, across Reddit, Twitter, LinkedIn, GitHub, and let the tool learn which tones work per platform', S4L fits.